CoChat
Sign Up
Cited by 125Remote-Sensing Image ClassificationRemote Sensing and Land UseDomain Adaptation and Few-Shot Learning

The DOI matches its Crossref record; 85 of the 85 references it cites were found in a registry or an index; 29 of the 55 sources its claims cite were checked for support.

Enabling country-scale land cover mapping with meter-resolution satellite imagery

Open access
Tong, Xin-Yi, Xia, Gui-Song, Zhu, Xiao Xiang·2023·ISPRS Journal of Photogrammetry and Remote Sensing
DOI 10.1016/j.isprsjprs.2022.12.011PMC PMC9939933
Open this paper in CoChat — free account

Abstract

High-resolution satellite images can provide abundant, detailed spatial information for land cover classification, which is particularly important for studying the complicated built environment. However, due to the complex land cover patterns, the costly training sample collections, and the severe distribution shifts of satellite imageries caused by, e.g., geographical differences or acquisition conditions, few studies have applied high-resolution images to land cover mapping in detailed categories at large scale. To fill this gap, we present a large-scale land cover dataset, Five-Billion-Pixels. It contains more than 5 billion labeled pixels of 150 high-resolution Gaofen-2 (4 m) satellite images, annotated in a 24-category system covering artificial-constructed, agricultural, and natural classes. In addition, we propose a deep-learning-based unsupervised domain adaptation approach that can transfer classification models trained on labeled dataset (referred to as the source domain) to unlabeled data (referred to as the target domain) for large-scale land cover mapping. Specifically, we introduce an end-to-end Siamese network employing dynamic pseudo-label assignment and class balancing strategy to perform adaptive domain joint learning. To validate the generalizability of our dataset and the proposed approach across different sensors and different geographical regions, we carry out land cover mapping on five megacities in China and six cities in other five Asian countries severally using: PlanetScope (3 m), Gaofen-1 (8 m), and Sentinel-2 (10 m) satellite images. Over a total study area of 60,000 km2, the experiments show promising results even though the input images are entirely unlabeled. The proposed approach, trained with the Five-Billion-Pixels dataset, enables high-quality and detailed land cover mapping across the whole country of China and some other Asian countries at meter-resolution.

Overview

AI snapshot

Written from the paper’s full text.

TL;DR

The paper introduces Five-Billion-Pixels, a 24-class, 4 m Gaofen-2 land cover dataset with more than 5 billion labeled pixels over 50,000 km² in China, and proposes a Siamese-network unsupervised domain adaptation method with dynamic pseudo-label assignment and class balancing for large-scale mapping. Experiments on five Chinese megacities and six additional Asian cities over 60,000 km² show the approach can map detailed land cover from unlabeled PlanetScope, Gaofen-1, and Sentinel-2 imagery.

Key findings

  • Five-Billion-Pixels contains more than 5 billion labeled pixels from 150 GF-2 images, covers over 50,000 km², and uses 24 categories spanning artificial-constructed, agricultural, and natural classes.
  • On the Five-Billion-Pixels benchmark, U-Net achieved OA 80.35%, mF1 57.34%, and mIOU 44.51%; DeepLabv3+ achieved OA 79.87%, mF1 54.84%, and mIOU 42.12%, outperforming object-oriented deep models and spectral–spatial methods.
  • For five Chinese megacities, the proposed approach reached sparse-label OA 74.08%, mF1 53.87%, mIOU 43.14%, and dense-label OA 81.28%, mF1 46.79%, mIOU 37.36%, outperforming DS-only and compared UDA methods.
  • For six additional Asian cities, the proposed approach reached sparse-label OA 81.14%, mF1 51.20%, mIOU 40.81%, and dense-label OA 80.85%, mF1 55.33%, mIOU 43.99%, with adversarial UDA methods showing severe negative adaptation.

Methods

  • The dataset was manually annotated in four phases—coarse labeling, fine labeling, fine checking, and spot checking—using GF-2 4 m multispectral imagery; 120 images were used for training and 30 for testing in the benchmark.
  • The UDA approach pre-trains U-Net on Five-Billion-Pixels, builds a Siamese network with shared parameters for source and target branches, assigns pseudo-labels to target pixels selected by lowest Shannon entropy, dynamically increases the number of selected pixels over training, and weights the joint loss by source-domain class distribution.
  • Land cover mapping was tested on five Chinese megacities using PlanetScope, Gaofen-1, and Sentinel-2 images and six additional Asian cities using Sentinel-2 images, with evaluation by OA, mF1, mIOU, and UA on sparse and dense labels.

Limitations

  • The paper states that obvious errors occur in forested mountains and at mosaic borders, that natural-class classification relies heavily on spectral information and can accumulate pseudo-label errors under spectral shifts, and that lake/river are heavily confused in the additional Asian cities while roads are less refined and discontinuous.
  • The paper states that pseudo-labels are difficult to capture contextual information, causing some local misclassifications such as park into pond and garden land in Chengdu, and that λ = 0.9 introduces too many pseudo-label errors while 150 epochs gives limited improvement over 100.
  • The text makes evident that the study covers only 11 cities rather than full country-scale mapping, that dense annotation was impossible for entire test images, and that comparison with Dynamic World and World Cover is only an approximate visual comparison due to different acquisition times.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1016/j.isprsjprs.2022.12.011

Cites

85 of 85 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

29 of 55 cited sources checked · 15 supported · 0 flagged · accuracy not yet measured

  • Findingcitation

    Five-Billion-Pixels extends the land cover dataset GID.

    “we propose the Five-Billion-Pixels dataset, which extends the land cover dataset GID (Tong et al., 2020, Yang et al., 2022).”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites Tong et al., 2020 source
    • cites Yang et al., 2022 source
  • Premisecitation

    A rigid alignment of two dispersive domains may further accumulate intra-domain variance.

    “In this context, a rigid alignment of two dispersive domains may further accumulate intra-domain variance (Tsai et al., 2018, Yan et al., 2019).”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites Tsai et al., 2018 source

      The Introduction says feature alignment across domains can be difficult for semantic segmentation because high-dimensional features encode diverse visual cues. Section 6.1 (GTA5) also reports that feature-space adaptation can fail to match source and target distributions effectively. Neither passage says rigid alignment of dispersive domains accumulates intra-domain variance; no page numbers are available in our copy. (LLM, conf=0.79)

      Experimental — accuracy not yet measured
    • cites Yan et al., 2019 source
  • Premisecitation

    Compared with conventional land cover classification methods, deep models can accurately characterize complex contextual information contained in high-resolution images.

    “Thus, compared with conventional land cover classification methods, deep models can accurately characterize complex contextual information contained in high-resolution images (Tong et al., 2020, Huang et al., 2018a, Zhang et al., 2019, Srivastava et al., 2019, Zhong et al., 2020).”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites Tong et al., 2020 source
    • cites Huang et al., 2018a

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Zhang et al., 2019 source
  • Premisecitation

    Current pseudo-labeling approaches empirically choose a confidence threshold or set a fixed proportion for sample collection, which is hard to guarantee the accuracy of pseudo-labels assigned by the prediction model to its unknown domain.

    “The current approaches are empirically choosing a confidence threshold (Xu et al., 2020, Peng et al., 2021) or setting a fixed proportion for sample collection (Tong et al., 2020, Zhang et al., 2021a, Yan et al., 2019, Wang et al., 2022), which is hard to guarantee the accuracy of pseudo-labels assigned by the prediction model to its unknown domain.”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites Xu et al., 2020

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Peng et al., 2021 source
    • cites Tong et al., 2020 source

      Section 3.2.2 says the method selects the top µ target samples in each category and acknowledges that pseudo-labels may be inaccurate. But µ is a fixed number, not a fixed proportion, and the method additionally checks pseudo-labels against labeled source samples before using them. The copy provides no page number for these passages. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
  • Premisecitation

    Deep models' performance strongly relies on the quality and quantity of training data.

    “their performance strongly relies on the quality and quantity of training data (LeCun et al., 2015, Xia et al., 2017, Ding et al., 2021)”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites LeCun et al., 2015 source
    • cites Xia et al., 2017 source

      Section 2.3 says that training large deep networks on aerial datasets with only hundreds or thousands of images can lead to overfitting and lower accuracy, supporting the importance of training-data quantity. The paper discusses dataset diversity and image relevance, but the cited passages do not establish a comparable dependence of model performance on training-data quality. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites Ding et al., 2021

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim. 41
  • It may state two separate claims as one. 55
  • The quote only partly states this claim. 1

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    An insufficient amount of data may lead to overfitting of the model.

    “An insufficient amount of data may lead to overfitting of the model”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites Long et al., 2021 source
    • cites Xiong et al., 2022

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • It may state two separate claims as one.

    Premisecitation

    Diverse imaging conditions lead to shifts in feature distributions of ground objects, which causes the classification method optimal for the source domain to drop drastically in performance on the target domain.

    “The attendant problem is that diverse imaging conditions will lead to shifts in feature distributions of ground objects, which causes the classification method optimal for certain annotated images (referred to as the source domain) to drop drastically in performance on newly acquired images (referred to as the target domain) (Tuia et al., 2016, Tong et al., 2020).”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites Tuia et al., 2016 source
    • cites Tong et al., 2020 source

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    The essential idea of discrepancy-based and adversarial-based UDA methods is to align the feature distributions of the source and target domains.

    “The essential idea of these two types of UDA methods is to align the feature distributions of the source and target domains (Sun and Saenko, 2016, Ganin et al., 2016, Wang et al., 2021).”
    Quote in the paperNot a substantial claim1 of 3 sources checked
    • cites Sun and Saenko, 2016 source
    • cites Ganin et al., 2016

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Wang et al., 2021 source

  • Probably not a substantial claim.

    Premisecitation

    Insufficient data diversity may lead to low generalization capability of the model.

    “insufficient data diversity may lead to low generalization capability of the model”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites Long et al., 2021 source
    • cites Xiong et al., 2022

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • It may state two separate claims as one.

    Premisecitation

    Most densely labeled land cover datasets with sub-meter to meter-spatial resolutions have geographical coverage areas below 10 km2 and are located in concentrated regions.

    “But most of them have geographical coverage areas below 10 km2 and are located in concentrated regions, such as ISPRS Potsdam (ISPRS-Contest, 2018), ISPRS Vaihingen (ISPRS-Contest, 2018), Zurich Summer (Volpi and Ferrari, 2015), RIT-18 (Kemker et al., 2018), and Zeebruges (Marcos et al., 2018).”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites ISPRS-Contest, 2018

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Volpi and Ferrari, 2015 source
    • cites Kemker et al., 2018 source

  • It may state two separate claims as one.

    Premisecitation

    World Cover and Dynamic World cover only basic land cover categories (11 and 9 classes) and have limited ability to depict urban environments.

    “Although they are highly accurate and real-time, they cover only basic land cover categories (11 and 9 classes) and have limited ability to depict urban environments.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites Zanaga et al., 2021

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Brown et al., 2022 source

      The Methods taxonomy (page unavailable) lists nine broad Dynamic World classes, including a single built-area class. Usage Notes says built area, as a mixed class, had lower accuracy in one derived-product assessment, but Technical Validation reports high agreement for built area in single-image predictions. The paper does not establish WorldCover’s 11-class count or assess its ability to depict urban environments, so it supports only the Dynamic World portion of the claim. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Premisecitation

    There are two general strategies for deep-learning-based land cover classification: object-oriented approaches based on DCNNs and semantic segmentation approaches based on end-to-end DCNNs.

    “There are two general strategies for deep-learning-based land cover classification: object-oriented approaches based on DCNNs and semantic segmentation approaches based on end-to-end DCNNs.”
    Quote in the paperNot a substantial claim2 of 3 sources checked
    • cites Huang et al., 2018a

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Zhang et al., 2019 source

      Section 2.3 describes an object-based CNN, but uses it for land-use rather than land-cover classification. Section 4.2 also discusses fully convolutional networks for semantic segmentation. The paper does not present these as the two general strategies for deep-learning-based land-cover classification. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites Srivastava et al., 2019 source

      The Introduction and Related Work distinguishes CNN-based patch classification from classifying all pixels in one forward pass, and the Methods contrasts urban-object classification with pixel-level semantic segmentation. It does not frame these as the two general strategies for deep-learning-based land-cover classification, or specify the claimed object-oriented DCNN versus end-to-end DCNN division. Page numbers are not available in our copy. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim. It may state two separate claims as one.

    Premisecitation

    Class balancing is a common strategy for training semantic segmentation models, but it is rarely used in UDA approaches because the category information in the target domain is unknown.

    “Class balancing is a common strategy for the training of semantic segmentation models (Long et al., 2015), but it is rarely used in UDA approaches because the category information in the target domain is unknown.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Long et al., 2015 source

  • Probably not a substantial claim.

    Premisecitation

    A Siamese network has two branches with identical architecture that share parameters during initialization and training, allowing it to learn information from two distributions simultaneously.

    “Siamese network has two branches, each of which have an input and an output (Zagoruyko and Komodakis, 2015). The two branches have an identical architecture and share the same parameters during both initialization and training, which allows the Siamese network to learn information from two distributions simultaneously, as presented in Fig. 7.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Zagoruyko and Komodakis, 2015 source

      Section 3.1 (page not provided) describes a Siamese network with two branches that have identical architecture and share weights; each branch takes one image patch as input and produces an output. It does not say that this arrangement lets the network learn from two distributions simultaneously, or discuss parameter sharing specifically during initialization. (LLM, conf=0.86)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it.

    Premisecitation

    The built environment and the quality of people's lives are under the combined effects of various land categories.

    “The built environment and the quality of people’s lives are under the combined effects of various land categories (Zhu et al., 2022)”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Zhu et al., 2022 source

      The Perspectives and discussions section says that how cities are physically built shapes quality of life, and the Experimental results section describes different built-up land classes and possible inequality in living conditions. The paper does not establish that various land categories have a combined effect on both the built environment and quality of life. (LLM, conf=0.82)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    U-Net achieves concatenation of each corresponding encoding and decoding stages using skip connection of feature maps, so that more raw information of the input image is retained and added to the decoding path.

    “It is worth noting that U-Net specially achieves concatenation of each corresponding encoding and decoding stages using skip connection of feature maps, as illustrated in Fig. 7, so that more raw information of the input image is retained and added to the decoding path.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Ronneberger et al., 2015 source

  • It may state two separate claims as one.

    Interpretationcitation

    U-Net's best performance on those categories is because recognition of these categories relies heavily on textural and spectral information, and U-Net retains more raw image information through its concatenation structure.

    “The recognition of these categories relies heavily on textural and spectral information. U-Net has stronger discriminative ability for them because it retains more raw image information through the concatenation structure (Ronneberger et al., 2015).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Ronneberger et al., 2015 source

      The Network Architecture section (p. 4) describes concatenating feature maps from the contracting path with upsampled features, and Experiments (pp. 6–7) reports strong segmentation results. The paper does not say that the cited categories depend on textural and spectral information, that concatenation retains more raw image information, or that this explains U-Net’s performance on those categories. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    Sentinel satellite data are utilized in the state-of-the-art global land cover mapping projects, ESA's World Cover and Google's Dynamic World.

    “Sentinel satellite data are utilized in the state-of-the-art global land cover mapping projects, ESA’s World Cover (Zanaga et al., 2021) and Google’s Dynamic World (Brown et al., 2022).”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites Zanaga et al., 2021

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Brown et al., 2022 source

  • It may state two separate claims as one.

    Findingnone given

    Agricultural and natural categories that are easier to classify occupy a larger total area in the additional six Asian cities, and there are fewer complex urban functional classes located in the test areas.

    “agricultural and natural categories that are easier to classify occupy a larger total area in these six cities, and there are fewer complex urban functional classes located in the test areas.”
    Quote in the paperSubstantial claim

  • It may state two separate claims as one.

    Findingown data

    Adversarial-based UDA methods demonstrate severe negative adaptation; they tend to classify complex areas into common categories when the distributions of the source and target domains are very different.

    “The adversarial-based UDA methods demonstrate severe negative adaptation; they tend to classify complex areas into common categories when the distributions of the source and target domains are very different.”
    Quote in the paperSubstantial claim

  • It may state two separate claims as one.

    Findingown data

    Fish pond along the coast of Guangzhou and paddy field in the suburbs of Wuhan are correctly identified by our approach.

    “Fish pond along the coast of Guangzhou and paddy field in the suburbs of Wuhan are correctly identified.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingcitation

    ResNet101 behaves better than GoogLeNet in overall and has significantly superior results on railway station, airport, shrub forest, natural meadow, and artificial meadow.

    “ResNet101 behaves better than GoogLeNet in overall and has significantly superior results on railway station, airport, shrub forest, natural meadow, and artificial meadow.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingown data

    DeepLabv3+ performs better on river than U-Net, and U-Net can recognize lake and irrigated field more accurately.

    “DeepLabv3＋ performs better on river than U-Net. And U-Net can recognize lake and irrigated field more accurately.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Findingown data

    The five Chinese megacities are sparsely labeled with 5.21×107 pixels and densely labeled with 7.77×106 pixels.

    “In total, the five Chinese megacities are sparsely labeled with 5.21×10⁷ pixels and densely labeled with 7.77×10⁶ pixels, and the percentage of each category in test areas is displayed in Table 2.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Findingreasoning

    The other six Asian cities are located in South, Southeast, and East Asia, respectively.

    “And the other six Asian cities are located in South, Southeast, and East Asia, respectively.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingreasoning

    The highest accuracy of Wuhan is reached when λ value is 0.7, while in other cities it is 0.5.

    “The highest accuracy of Wuhan is reached when λ value is 0.7, while in other cities it is 0.5.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingreasoning

    U-Net behaves best on industrial area, urban residential, rural residential, different agriculture, and different forest classes.

    “U-Net behaves best on industrial area, urban residential, rural residential, different agriculture, and different forest classes.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingreasoning

    Regardless of the method type, there are high performance discrepancies of different categories; all methods behave poorly on square, overpass, railway station, airport, garden land, park, and pond.

    “Another issue worth noting is that regardless of the method type, there are high performance discrepancies of different categories. For instance, all methods behave poorly on square, overpass, railway station, airport, garden land, park, and pond.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingown data

    GoogLeNet and ResNet101 fail in extracting road and misclassify paddy field into fish pond.

    “GoogLeNet and ResNet101 fail in extracting road and misclassify paddy field into fish pond.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingreasoning

    There are significant differences between data sources and distribution shifts between images within each city.

    “As a result, not only are there significant differences between data sources, but there are also distribution shifts between images within each city, as displayed in Fig. 4.”
    Quote in the paperSubstantial claim

  • It may state two separate claims as one.

    Findingreasoning

    Five-Billion-Pixels is mainly collected from human activity areas (cities, villages, cultivated lands, and mountainous areas around cities), and its category system covers all land categories except mangroves, tundra and permanent ice.

    “As Five-Billion-Pixels are mainly collected from human activity areas (cities, villages, cultivated lands, and mountainous areas around cities), the category system covers all land categories except mangroves, tundra and permanent ice.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingown data

    DeepLabv3+ can segment more complete and smooth airport, bare land, and stadium, while U-Net results contain a lot of noise.

    “In columns 1 to 3, DeepLabv3＋ can segment more complete and smooth airport, bare land, and stadium, respectively. And the results of U-Net contain a lot of noise.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingown data

    The approach improves the results of paddy field in Bangkok, railway station in Delhi, irrigated field in Naypyidaw, arbor forest in Seoul, industrial area in Tokyo, and airport in Yangon.

    “our approach improves the results of paddy field in Bangkok, railway station in Delhi, irrigated field in Naypyidaw, arbor forest in Seoul, industrial area in Tokyo, and airport in Yangon.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingown data

    MLP+Fusion can identify some water areas, RF+Fusion can identify some water, built-up, and paddy field areas, but the rest of the map is heavily confused.

    “MLP＋Fusion can identify some water areas, RF＋Fusion can identify some water, built-up, and paddy field areas, but the rest of the map is heavily confused.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingnone given

    Our approach is able to distinguish industrial area, urban residential in city center, rural residential dispersed in suburb, transportation networks, and river systems for each of five Chinese megacities, despite images from different sensors and no annotated information.

    “Although images from different sensors are utilized, and there is no annotated information on them, our approach is able to distinguish industrial area, urban residential in city center, rural residential dispersed in suburb, transportation networks, and river systems for each city.”
    Quote in the paperSubstantial claim

  • It may state two separate claims as one.

    Findingown data

    DeepLabv3+ and U-Net can segment clear road networks and different built-up areas.

    “In contrast, DeepLabv3＋ and U-Net can segment clear road networks and different built-up areas.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Findingown data

    The category system of Five-Billion-Pixels contains a number of land use classes subdivided from land cover classes in GB/T 21010-2017.

    “Notably, the category system contains a number of land use classes subdivided from land cover classes in GB/T 21010-2017, including: stadium and square from public service land; road, overpass, railway station, and airport from transportation land; park and artificial meadow from artificial non-agricultural vegetated areas.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Findingown data

    The additional Asian cities are sparsely labeled with a total of 2.40×106 pixels and densely labeled with a total of 1.29×106 pixels.

    “The additional Asian cities are sparsely labeled with a total of 2.40×10⁶ pixels and densely labeled with a total of 1.29×10⁶ pixels, and the percentage of each category in test areas is listed in Table 3.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Findingnone given

    The Chinese megacities are separately located in the eastern, western, northern, southern, and central regions of China.

    “The Chinese megacities cover a total geographical area of 53,088 km2 and are separately located in the eastern, western, northern, southern, and central regions of China.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingreasoning

    In Delhi, Naypyidaw, and Seoul, comparison methods classify built-up areas into irrigated field, and in Tokyo, high-rise residential and industrial area are misclassified as low-rise residential.

    “in Delhi, Naypyidaw, and Seoul, the comparison methods classify built-up areas into irrigated field, and in Tokyo, high-rise residential and industrial area are misclassified as low-rise residential.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingreasoning

    DeepLabv3+ tends to identify ground objects as independent "instances", whereas U-Net struggles to aggregate land information into homogeneous segmented regions.

    “DeepLabv3＋ tends to identify ground objects as independent “instances”, whereas it seems difficult for U-Net to aggregate land information into homogeneous segmented regions.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Findingnone given

    The Chinese megacities cover a total geographical area of 53,088 km2.

    “The Chinese megacities cover a total geographical area of 53,088 km2 and are separately located in the eastern, western, northern, southern, and central regions of China.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingreasoning

    Results on dense label are generally better than on sparse label in OA, while sparse label outperforms dense label in mF1 and mIOU.

    “It can be seen that the results on dense label are generally better than that on sparse label in OA, while sparse label outperforms dense label in mF1 and mIOU.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingreasoning

    Obvious errors occur in the forested mountains and at the mosaic borders; small areas of mountains around Beijing are misclassified as water bodies, and some areas of mountains around Chengdu are misclassified as irrigated field.

    “Obvious errors occur in the forested mountains and at the mosaic borders. Small areas of mountains around Beijing are misclassified as water bodies, and some areas of mountains around Chengdu are misclassified as irrigated field.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingown data

    U-Net shows superior performance in columns 4 to 5, correctly classifying irrigated field, which are partially misclassified by DeepLabv3+ as garden land and road, which cannot be clearly extracted by DeepLabv3+ from dense built-up areas.

    “U-Net shows superior performance in columns 4 to 5, where it correctly classifies irrigated field, which are partially misclassified by DeepLabv3＋ as garden land and road, which cannot be clearly extracted by DeepLabv3＋ from dense built-up areas.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Findingreasoning

    New high-rise residential buildings are mixed with old ones, which have the same appearance as rural settlements, creating a special urban landscape in China.

    “new high-rise residential buildings are mixed with old ones, which have the same appearance as rural settlements, creating a special urban landscape.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Findingnone given

    The Five-Billion-Pixels dataset covers a large amount of irrigated field reclaimed in the mountains of northwestern China, and their spectral and texture features may be similar to those of mountainous arbor forest in winter.

    “the Five-Billion-Pixels dataset covers a large amount of irrigated field reclaimed in the mountains of northwestern China, and their spectral and texture features may be similar to those of mountainous arbor forest in winter.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Interpretationnone given

    The category system of Five-Billion-Pixels covers artificial-constructed, agricultural, and natural classes, more closely resembling the distributions of ground objects in the real world.

    “This category system covers artificial-constructed, agricultural, and natural classes, more closely resembling the distributions of ground objects in the real world.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Study design and conductnone given

    The labeled areas of Five-Billion-Pixels are double-checked and ensured to be correct.

    “The labeled areas are double-checked and ensured to be correct.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper may give no reason why it holds. It may state two separate claims as one.

    Premisenone given

    The model always tends to select more easy samples, which may exacerbate the category imbalance.

    “Furthermore, the model always tends to select more easy samples, which may exacerbate the category imbalance.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. It may be too vague to check against evidence. The paper may give no reason why it holds.

    Premisenone given

    Dense annotation for each newly captured image is not realistic.

    “However, dense annotation for each newly captured image is not realistic.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim.

    Premisecitation

    Deep Convolutional Neural Networks can adaptively approximate the relationship between image information and land information through multi-layer transformations.

    “Deep Convolutional Neural Networks (DCNNs) are able to adaptively approximate the relationship between image information and land information through multi-layer transformations (Zhu et al., 2017).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Zhu et al., 2017 source

  • It may state two separate claims as one.

    Premisecitation

    The detailed information in high-resolution images brings about much more complicated land structures and patterns, which leads to great challenges in land cover classification.

    “But at the same time, the detailed information brings about much more complicated land structures and patterns (Yuan et al., 2020), which leads to great challenges in land cover classification with high-resolution images.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Yuan et al., 2020 source

  • Probably not a substantial claim. The paper may give no reason why it holds.

    Premisereasoning

    Few studies have applied high-resolution images to land cover mapping in detailed categories at large scale.

    “few studies have applied high-resolution images to land cover mapping in detailed categories at large scale.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim.

    Premisereasoning

    The quality of pseudo-labels depends on the way they are selected.

    “However, the quality of pseudo-labels depends on the way they are selected.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim. The paper may give no reason why it holds. It may state two separate claims as one.

    Interpretationnone given

    The category system of Five-Billion-Pixels covers artificial-constructed, agricultural, and natural classes, which well-reflects the distribution of real-world ground objects and can widely benefit land-cover-related studies.

    “Its category system covers artificial-constructed, agricultural, and natural classes, which well-reflects the distribution of real-world ground objects and can widely benefit land-cover-related studies.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Interpretationreasoning

    The distinctive characteristic of those categories is that they are composed of multiple basic ground cover types, which causes them to be easily misclassified into other categories.

    “In addition, the distinctive characteristic of these categories is that they are composed of multiple basic ground cover types; for example, railway station contains multiple tracks and stadium-like building roof, park includes grass and woods, and airport contains roads and lawns, which causes them to be easily misclassified into other categories.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Interpretationreasoning

    The performance advantages of deep learning, especially of semantic segmentation models, demonstrate the importance of large-scale, pixel-wise annotated datasets for advancing land cover classification research.

    “The performance advantages of deep learning, especially of semantic segmentation models, demonstrate the importance of large-scale, pixel-wise annotated datasets for advancing land cover classification research.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. It may state two separate claims as one.

    Premisenone given

    Object-oriented approaches and semantic segmentation approaches have different requirements for training data: object-oriented approaches allow each input sample to have one label, while semantic segmentation approaches require a label for each pixel.

    “Since object-oriented approaches and semantic segmentation approaches have different requirements for training data, i.e., object-oriented approaches only allow each input sample to have one label, while semantic segmentation approaches require a label for each pixel of input sample, we prepare different training data for them.”
    Quote in the paperSubstantial claim

  • It may state two separate claims as one.

    Interpretationreasoning

    When λ is set to 0.9, 90% pixels in the target domain are involved in the last training epoch, too many errors in pseudo-labels, thus leading to an obvious decrease in the accuracy of each city.

    “when λ is set to 0.9, 90% pixels in the target domain are involved in the last training epoch, too many errors in pseudo-labels, thus, lead to an obvious decrease in the accuracy of each city.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Interpretationreasoning

    The discontinuity of road lines is due to feature distribution differences caused by diverse architectural styles and city landscapes in different countries.

    “This is due to the feature distribution differences caused by the diverse architectural styles and city landscapes in different countries.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    The latest semantic segmentation models typically consist of two main paths: an encoder path that reduces feature map size and captures higher-level information, and a decoder path that recovers spatial resolution and object boundaries.

    “The latest semantic segmentation models typically consist of two main paths: the encoder path that gradually reduces feature map size and captures higher-level information, and the decoder path that gradually recovers spatial resolution as well as clear object boundaries (Chen et al., 2018).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Chen et al., 2018

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • It may state two separate claims as one.

    Interpretationcitation

    ResNet101's superior performance on those categories is due to its residual connection structure that enables combination of different levels of features, allowing it to learn low-level features to distinguish natural classes and high-level features to identify artificial buildings with complex structures.

    “Owing to the residual connection structure (He et al., 2016) that enables the combination of different levels of features, ResNet101 can learn low-level features to distinguish natural classes as well as high-level features to identify artificial buildings with complex structures.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites He et al., 2016

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim. It may be too vague to check against evidence. The paper may give no reason why it holds.

    Interpretationnone given

    Compared to discrepancy-based and adversarial-based UDA methods that force the two distributions to be aligned in feature space, pseudo-labeling is more flexible and potentially more dependable for complicated real-world situations.

    “Compared to discrepancy-based and adversarial-based UDA methods, which force the two distributions to be aligned in feature space, pseudo-labeling is more flexible and potentially more dependable for complicated real-world situations.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Interpretationreasoning

    Semantic segmentation models outperform object-oriented methods because they can capture contextual information over larger areas and simultaneously maintain more accurate edges for ground objects by assigning labels to each pixel.

    “This is because semantic segmentation models can capture contextual information over larger areas and simultaneously maintain more accurate edges for ground objects by assigning labels to each pixel.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim. The paper may give no reason why it holds.

    Interpretationnone given

    If there are errors in pseudo-labels, a small number of mistakes may eventually lead to a relatively large bias during iterative training.

    “If there are errors in pseudo-labels, a small number of mistakes may eventually lead to a relatively large bias during the iterative training.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Interpretationcitation

    DeepLabv3+'s best results on those categories are due to atrous convolution and spatial pyramid pooling, which allow it to capture multi-level contextual information.

    “Due to the atrous convolution (Chen et al., 2017b) and spatial pyramid pooling (Chen et al., 2018) adopted in DeepLabv3＋, it can capture multi-level contextual information for these categories.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Chen et al., 2017b

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Chen et al., 2018

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim. The paper may give no reason why it holds.

    Interpretationnone given

    The proposed UDA approach can softly correct the domain shifts by adaptively learning the distribution of unlabeled data.

    “we propose a UDA approach that can softly correct the domain shifts by adaptively learning the distribution of unlabeled data.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Interpretationreasoning

    Using only a very small number of pseudo-labels in the initial iterations of domain joint learning leads to two problems: DT provides very few training samples at the beginning, and the samples selected from DT may be extremely homogeneous.

    “To avoid introducing incorrect category information in the training, only a very small number of pseudo-labels are used in the initial iterations of domain joint learning, which leads to two problems: (1) DT can only provide very few training samples at the beginning; (2) the samples selected from DT may be extremely homogeneous.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Study design and conductnone given

    These annotations are only used for accuracy assessment and not for model training.

    “Note that these annotations are only used for accuracy assessment and not for model training.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Interpretationreasoning

    The greater performance gap in the additional six Asian cities is because the feature distributions of the source and target domains are more disparate, making it harder for adversarial-based approaches to find a suitable match and introducing more intra-domain variance by rigid alignment.

    “This is because in this scenario, the feature distributions of the source and target domains are more disparate, it is harder for the adversarial-based approaches to find a suitable match between the two distributions, and more intra-domain variance is introduced into the model by the rigid alignment.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim. It may state two separate claims as one.

    Interpretationreasoning

    The network may give reliable predictions for only a small number of pixels when it is not adapted to DT, and as it gradually learns the distribution of DT, it can make reliable predictions for an increasing number of pixels.

    “An intuitive interpretation of this design is that the network may give reliable predictions for only a small number of pixels when it is not adapted to DT, and as the network gradually learns the distribution of DT, it can make reliable predictions for an increasing number of pixels.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisenone given

    Unlike artificial-constructed categories, the classification of natural classes relies more on spectral information.

    “Unlike artificial-constructed categories, the classification of natural classes relies more on spectral information.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Interpretationreasoning

    The higher optimal λ in Wuhan may be because the sensors of GF-1 and GF-2 have more similar imaging processes, resulting in closer distributions of the source and target domains, and the quality of the pseudo-label can be guaranteed to some extent even if 70% pixels are selected at the end of training, whereas using more pseudo-labels in other cities would introduce more errors.

    “This may be because the sensors of GF-1 and GF-2 have more similar imaging processes, resulting in closer distributions of the source and target domains, and the quality of the pseudo-label can be guaranteed to some extent even if 70% pixels are selected at the end of training, whereas using more pseudo-labels in other cities would introduce more errors.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: GF-1 and GF-2 have more similar imaging processes, resulting in closer distributions of the source and target domains.

  • It may state two separate claims as one.

    Premisecitation

    Shannon Entropy is used as an indicator to quantify the confidence of each image pixel, with lower information entropy representing higher classification confidence.

    “Shannon Entropy (Xu et al., 2020) is employed as our indicator to quantify the confidence of each image pixel. Lower information entropy represents higher classification confidence.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Xu et al., 2020

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • It may state two separate claims as one.

    Interpretationreasoning

    The poor performance on those categories is due to two factors: they represent small percentages in the Five-Billion-Pixels dataset, and they are inherently easier to confuse with other categories.

    “This is due to two factors, first, these categories represent small percentages in the Five-Billion-Pixels dataset, and second, they are inherently easier to confuse with other categories.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Interpretationreasoning

    When spectral shifts are particularly significant, pseudo-labels inevitably contain errors, which will continuously accumulate in iterative domain adaptation learning.

    “When the spectral shifts are particularly significant, pseudo-labels inevitably contain errors, which will continuously accumulate in iterative domain adaptation learning.”
    Quote in the paperSubstantial claim

  • It may state two separate claims as one.

    Interpretationreasoning

    In domain joint learning, the DCNNs model is likely to assign irrigated field pseudo-labels to a small number of arbor forest samples in DT, and these errors will further accumulate in continuous iterative training, eventually leading to misclassification in the land cover mapping results.

    “the DCNNs model is likely to assign irrigated field pseudo-labels to a small number of arbor forest samples in DT. And these errors will further accumulate in continuous iterative training, eventually leading to misclassification in the land cover mapping results.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Features of forest and cropland are largely influenced by geographical location and seasonal changes.

  • Probably not a substantial claim. The paper may give no reason why it holds.

    Interpretationnone given

    The approach is proven to be generalizable across images captured by different sensors, especially for ST-2 images that are free and open.

    “our approach is proven to be generalizable across images captured by different sensors, especially for ST-2 images that are free and open.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper may give no reason why it holds. It may state two separate claims as one.

    Interpretationreasoning

    This problem is caused by the inherent constraints of UDA and the intrinsic properties of vegetation categories.

    “Since this problem is caused by the inherent constraints of UDA and the intrinsic properties of vegetation categories”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Interpretationown data

    Owing to the extensive, fine, and accurate annotating for Five-Billion-Pixels, the results present clear transportation networks and river systems, as well as different agricultural and urban functional areas.

    “owing to our extensive, fine, and accurate annotating for Five-Billion-Pixels, our results present clear transportation networks and river systems, as well as different agricultural and urban functional areas.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Interpretationreasoning

    Although only 11 cities were studied, the approach can be easily generalized to other cities, towns, and villages throughout China and even other Asian countries.

    “Although we studied only 11 cities in this paper, our approach can be easily generalized to other cities, towns, and villages throughout China and even other Asian countries.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim. The authors may not put it forward as true. It may be too vague to check against evidence. The paper’s conclusions may not depend on it.

    Interpretationreasoning

    For other countries and regions with very different land distributions and land category systems than China, it has potential to perform land cover mapping based on Five-Billion-Pixels and semi-supervised domain adaptation with sparse annotations in the form of patches or polygons.

    “for other countries and regions with very different land distributions and land category systems than China, it has potential to perform land cover mapping based on Five-Billion-Pixels and, for example, semi-supervised domain adaptation with sparse annotations in the form of patches or polygons.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. The quote only partly states this claim.

    Study design and conductown data

    Due to the different acquisition time of the original images, the comparison with Dynamic World and World Cover can only be an approximate visual comparison.

    “Due to the different acquisition time of the original images, we can only make an approximate visual comparison.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim. It may be too vague to check against evidence. It may state two separate claims as one.

    Interpretationown data

    The work has the potential to be extended to land cover mapping at country-scale and to contribute to various applications involving land cover information.

    “our work has the potential to be extended to land cover mapping at country-scale and to contribute to various applications involving land cover information.”
    Quote in the paperNot a substantial claim
80 of 80

  • The paper relies on this without support.

    The 10% sample examined in the final spot checking is representative of the entire Five-Billion-Pixels dataset.

    “In the final round of spot checking, 10% samples of Five-Billion-Pixels are examined and no obvious errors are observed.”

  • The paper relies on this without support.

    The effective spatial resolution of PS is 3.7–4.1 m.

    “Since the 3 m resolution of PS is obtained by resampling the raw data and its effective spatial resolution is 3.7−4.1 m, we use the original image resolution of GF-2 to adapt it.”

  • The paper relies on this without support.

    The Five-Billion-Pixels dataset is representative of the target cities' land cover distributions.

    “To adapt deep models to satellite images with different resolutions, we construct a multi-scale source domain using Five-Billion-Pixels.”

  • The paper relies on this without support.

    Land cover mapping at large scale in high resolution remains a challenging task.

    “Yet land cover mapping on a large-scale in high-resolution remains a challenging task.”

  • The paper relies on this without support.

    Recognition of heterogeneous ground objects in high-resolution satellite images is difficult.

    “the recognition of heterogeneous ground objects in high-resolution satellite images is quite difficult.”

  • The paper relies on this without support.

    Different land cover categories depend on different types of information for identification: urban functional buildings rely more on structure and shape features, dense built-up areas require spatial relationships, and agricultural and natural categories require texture and spectral information.

    “classifying independent urban functional buildings relies more on structure and shape features, distinguishing dense built-up areas requires spatial relationships, and the identification of agricultural and natural categories cannot be done without texture and spectral information.”

  • The paper relies on this without support.

    ST-2 has lower spatial resolution than PS and GF-1, and cannot present very fine edges in both ground truth and results.

    “the spatial resolution of ST-2 is lower than those of PS and GF-1, and it cannot present very fine edges in both ground truth and results.”

  • The paper relies on this without support.

    GF-1 and GF-2 have more similar imaging processes, resulting in closer distributions of the source and target domains.

    “the sensors of GF-1 and GF-2 have more similar imaging processes, resulting in closer distributions of the source and target domains”

  • The paper relies on this without support.

    Features of forest and cropland are largely influenced by geographical location and seasonal changes.

    “their features are largely influenced by geographical location and seasonal changes.”

  • The paper relies on this without support.

    Rapid urban development in China causes new high-rise residential buildings to be mixed with old ones, which have the same appearance as rural settlements.

    “because of the rapid urban development in China, new high-rise residential buildings are mixed with old ones, which have the same appearance as rural settlements”
10 of 10

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Entropy is calculated from feature values rather than probabilities

    Minor

    The stated entropy formula takes the logarithm of feature values, which are specified as arbitrary real numbers; it can therefore be undefined for negative values and is not necessarily between 0 and 1. The paper obtains category probabilities only afterward by applying softmax.

  • Table 7 reports an impossible combination of mean F1 and mean IOU

    Minor

    For each category, F1 = 2·IOU/(1+IOU). Consequently, a mean IOU of 19.21% permits a mean F1 of at most 32.23%, but the AdaptSeg+DPA row reports 32.35%; rounding cannot account for the difference.

2 of 2

Citations

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. Lee, D.-H., et al., 2013. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In: International Conference on Machine Learning Workshops, Vol. 3. p. 896.

Cited by

Sorted by citations

Related

Metadata

Authors
Tong, X.-Y., Xia, G.-S., & Zhu, X. X.
Affiliations
  • Wuhan University (CN)
  • Technical University of Munich (DE)
Corresponding author
Zhu, Xiao Xiang
ORCID
Venue
ISPRS Journal of Photogrammetry and Remote Sensing
Publisher
Elsevier BV
Published
2023-02
Volume
196
Pages
178-196
ISSN
0924-2716 (ISSN-L)
DOI
10.1016/j.isprsjprs.2022.12.011
PMC
PMC9939933
OpenAlex
W4313908836
Semantic Scholar
5f23ad23dbe17acba04e07fa5221576d5bb11a58
Cited by
125
References
85
Citations by year
2026: 26 · 2025: 55 · 2024: 34 · 2023: 10
FWCI
9.31
Type
journal article
Language
en
Open access
Licence
Locations
  • ISPRS Journal of Photogrammetry and Remote Sensing (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • elib (German Aerospace Center) (acceptedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • European Commission (HORIZON2020)
  • Deutsche Forschungsgemeinschaft (519016653, ZH 498/18-1)
  • Bundesministerium für Bildung und Forschung (01DD20001)
  • Bundesministerium für Wirtschaft und Technologie (W2-W3-100, 50EE2201C)
  • Helmholtz Artificial Intelligence Cooperation Unit (ZT-I-PF-5-01)
  • H2020 European Research Council (714087)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Remote-Sensing Image Classification
  • Social Sciences › Geography, Planning and Development › Remote Sensing and Land Use
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
Keywords
land cover mapping, high-resolution satellite imagery, unsupervised domain adaptation, Siamese network, pseudo-labeling, data balancing, Gaofen-2, PlanetScope, Sentinel-2, Gaofen-1, cross-sensor generalization, cross-regional generalization, built environment

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-08, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-08 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review